Building A Personal Deep Learning Machine

A worrier cannot fight longer on a rented sword.

I have started working on deep learning in late 2014. At that time datasets were small, very few deep learning frameworks were available. Back then Caffe and H2O were available as only practical frameworks. Caffe was purely on C side and H2O was available in R and Python. Caffe had buggy support for CNN and H2O was only supporting Feed Forward Network (FFN) along with machine learning techniques like SVM, GBM, and GLM. I managed to complete my thesis utilizing Caffe and H2O on predicting alignments of protein in 3D spaces.

In the next 5 year time has changed a lot. Python became the language of choice for deep learning with Tensorflow, MXNet, Keras, and PyTorch started porting to python. Data set grew quickly and some precious ones became openly available. An explosion happened in the algorithm variants in the deep learning domain and quickly stable support for CNN, RNN, RL, and GAN made available by the combined effort of open source community and research-based companies.

For the very last few months, I was been extensively using Google Colab for many of my experiments related to deep learning. I feel thankful to Google for showing the courage to make the available GPU for free. I respect the courage but one cannot have a bigger dataset and cannot run an experiment for more than 12 hours. Besides this, with Google Colab, I cannot have a GUI where I can run simulators like Carla. Quickly the hunger for power grew and I felt that Google Colab is not enough and there has to be something personal. I started weighing my options.

I could go for cloud computing but I had following limitations of cloud in mind

• As happened with Google Colab, I will be having a problem with GUI access and keeping the large dataset.
• Cloud computing with huge data, good GPU devices, and bandwidth turned out to be a relatively costlier option (on personal capacity).

The other way around was buying a high-end laptop or a desktop. I was looking for more compute than Nvidia 1060 and 1070 and having it in a laptop was difficult.

The laptop is a good choice if you really want to carry it everywhere you go but there are certain limitations :

1. Usually OEM like MSI, Dell, and HP do not provide a laptop with full-size cards like 1080 Ti. Usually, the laptop comes with 6-8GB GPU VRAM and costs around $3000 USD
2. The GPU is very good at consuming the battery so whenever you run a job on GPU your laptop must be plugged to the main power.

On the other hand, the desktop PC will be even bulkier but you can put as much power you want. Meanwhile, when I fixed my direction for desktop, I also considered ROG Mothership and Corsair One as the option. These two are a really good option to opt but these builds, unfortunately, were not available in India.

I wanted a small yet powerful build, I had fixed some of the variables and then started looking for parts accordingly

1. I wanted a small build which can support one full-size GPU card like 1080Ti
2. I wanted my build to fit into a trolly bag so that I can carry it very easily.

Considering above mentioned priorities I started ordering the components from various online retails and local retailers. One thing is sure, All online retails are costlier compared to the local retailers in these kinds of special computer parts. One may find such local retailers by visiting a local classified portal such as OLX, Quickr, or Justdial. I found one distributor specifically dealing with gaming rigs. The difference in price between this distributor and online retails will be evident as you see below given list of components along with the price.

1. GPU - 1080Ti Founders edition - 90,000 INR
2. SSD 256 GB : ADATA XPG GAMMIX S5 256GB M.2 -- 5,084.75 INR
3. Motherboard: ASROCK Z370M PRO4 M-ATX MOTHER -- 9,745.76 INR
4. RAM: HYPERX PREDATOR DDR4 8GB 3000M * 2 -- 10,169.50 INR
5. CPU: INTEL CORE i7 8700 3.2 GHZ PRO -- 25,000.00 INR
6. Cabinet: NZXT H400 BLACK M-ATX CABINET -- 7,627.12 INR
7. CPU Liquid Cooler: NZXT X42 KRAKEN 120MM LIQUID C -- 9,322.03 INR
8. SMPS: TT TOUGH POWER GRAND 650W GOLD -- 8,474.58 INR
9. SSD: INTEL 512GB SSD -- 6,355.93 INR & more drive as per need

Now comes the assembling parts, all the cables are so designed that only a dumb can make a mistake. But I admit its equally confusing unless you go through all the manuals provided along with products. Here is where I made a mistake. I started building without reading the manual and halted the build with all wires running around.

On the next day early morning, I again started with building the rig. figured out how to wire all chassis fans, attached the liquid cooler to CPU, and arranged radiator. Attached all the cables to the motherboard and SMPS. With my heart beating very fast, keeping all the cables hanging around I first connected I to monitor and tested If it's working. Finally, I saw the motherboard logo and boot screen. After 6+ Hr of frustrating work, I finally felt relaxed. Then at the end attached GPU to the PCI slot. Thereafter I quickly arranged all the cables, thanks to tools and manuals, Finally, My PC was complete.

NOW: I built this rig one a year back just found time to post it now. By this time I have Joined Nvidia as Deep Learning - Data Scientist, a dream come true experience. I have transferred this build multiple times (it fits well in medium size trolly) in flight and via road transport still, it works like charm. Now I have replaced 1080ti with Nvidia RTX - 8000, a more powerful GPU and added more RAM sticks. I hope this build experience will help you lot in building your first RIG>